Mutually Informed Correlation Coefficient (MICC) - a New Filter Based Feature Selection Method

Abstract

Feature selection (FS) is a well-explored domain of data pre-processing and information theory. It is the process of selecting important features from a high-dimensional feature vectors possibly having many redundant and/or non-informative features. In this paper, we have proposed a score-based filter FS approach known as Mutually Informed Correlation Coefficient (MICC) by combining two popular statistical dependence measures namely Mutual Information (MI) and Pearson Correlation Coefficient (PCC). We have evaluated MICC on different variations of Local Binary Pattern (LBP) based feature vectors used for classifying the components of handwritten document images as text or non-text. We have compared the results with some popular filter methods namely Gini Index, T-test, ReliefF, along with MI and PCC individually. The results and corresponding comparisons show that our proposed method not only does FS efficiently but also enhances the recognition accuracy of the said classification problem. The code of the proposed algorithm can be found in this link: MICC.